The ability to automatically interpret images is one of the major challenges for modern computer vision. This paper describes the process by which images may be classified using neural network techniques. These techniques are quantitatively evaluated using the Bristol Image Database, a large set of high-quality colour images of outdoor scenes, with a known ground-truth labelling. Once an image has been automatically segmented into regions, the regions are classified by training a multi-layer perceptron with a compact set of features extracted from each region. The features used include shape, texture, colour and context.